Ear recognition based on 3D keypoint matching

This paper proposes a novel ear recognition approach based on 3D keypoint matching. At first, the 3D keypoints are detected using the shape index image and the scale space theory. Then two principal orientations are assigned and the normalized local range image is obtained, which can provide invariance to 3D rotation and transformation for the following local descriptor construction. Finally, we construct the 3D CS-LBP features and use a coarse to fine strategy for 3D keypoint matching. The number of the matching points and their average EMD distances are used for 3D ear recognition. The proposed approach can reduce the amount of 3D data by 3D keypoint detection and local feature construction, and it doesn't need any expensive preprocessing steps. Compared with existing 2D or 3D LBP operators, the proposed 3D CS-LBP operator can not only remain the 3D LBP's powerful ability to describe the 3D structure information, but also reduce the histogram size and enhance its robustness to noise. Extensive experiments have performed to valid the efficiency of the proposed approach.